Automatic labeling method for unlabeled data of point clouds
An automatic labeling method for assigning labels to unlabeled point clouds among a set of labeled and unlabeled point clouds includes preparing an initial machine learning classification model, selecting a labeled point cloud for each of the unlabeled point clouds based on similarities between a feature vector of each of the unlabeled point clouds output through the model and feature vectors of the labeled point clouds output through the model and assigning a cluster label to each of the unlabeled point clouds based on a label of the selected labeled point cloud, assigning pseudo labels to the unlabeled point clouds to which the cluster labels are assigned based on a confidence score obtained through the model, and updating the model by training the model with the labeled point clouds and the unlabeled point clouds to which the pseudo labels are assigned.
1. An automatic labeling method for automatically assigning labels to unlabeled point clouds among a set of labeled and unlabeled point clouds, the automatic labeling method comprising:
(a) preparing an initial machine learning classification model;
(b) selecting a labeled point cloud for each of the unlabeled point clouds based on similarities between a feature vector of each of the unlabeled point clouds output through the model and feature vectors of the labeled point clouds output through the model and assigning a cluster label to each of the unlabeled point clouds based on a label of the selected labeled point cloud;
(c) assigning pseudo labels to the unlabeled point clouds to which the cluster labels are assigned, wherein each pseudo label is assigned based on a confidence score obtained through the model; and
(d) updating the model by training the model with the labeled point clouds and the unlabeled point clouds to which the pseudo labels are assigned, wherein, steps (b) through (d) are performed iteratively after step (d).
2. The automatic labeling method of claim 1 , wherein step (c) comprises:
selecting a part of the unlabeled point clouds based on silhouette scores thereof obtained through the model; and
assigning the pseudo labels, wherein each pseudo label is assigned based on a prediction result probability value obtained through the model for each point cloud of the selected part.
3. The automatic labeling method of claim 2 , wherein each pseudo label is assigned with a temporary label or the cluster label according to whether the temporary label and the cluster label are the same, and wherein the temporary label is determined based on the prediction result probability value obtained through the model.
4. The automatic labeling method of claim 3 , wherein the unlabeled point clouds to which pseudo labels are not assigned in the selected part remain unlabeled.
5. The automatic labeling method of claim 2 , wherein step (c) further comprises assigning pseudo labels to the unlabeled point clouds that are not included in the selected part, wherein each pseudo label is assigned based on the prediction result probability value obtained through the model.
6. The automatic labeling method of claim 5 , wherein each pseudo label is assigned to the unlabeled point cloud if a temporary label corresponds to a class that is not in the feature vectors of the unlabeled point clouds of the selected part, and wherein the temporary label is determined based on the prediction result probability value obtained through the model for the corresponding unlabeled point cloud.
7. The automatic labeling method of claim 6 , wherein the unlabeled point clouds to which pseudo labels are not assigned among the unlabeled point clouds not included in the selected part remain unlabeled.
8. The automatic labeling method of claim 1 , wherein step (a) comprises, after unsupervised learning that uses the unlabeled point clouds, constructing the model through supervised learning that uses the labeled point clouds.
9. The automatic labeling method of claim 1 , wherein the iterative performing is terminated when the pseudo labels are no longer assigned in step (c).
10. The automatic labeling method of claim 1 , wherein step (b) comprises:
selecting a preset number of labeled point clouds for each of the unlabeled point clouds based on the similarities; and
assigning, as the cluster label, a label with a highest frequency among labels of the selected labeled point clouds.
11. The automatic labeling method of claim 10 , wherein the preset number is 20 or greater.
12. An automatic labeling method for automatically assigning labels to unlabeled point clouds among a set of labeled and unlabeled point clouds, the automatic labeling method comprising:
preparing an initial machine learning classification model by unsupervised learning followed by supervised learning;
selecting a labeled point cloud for each of the unlabeled point clouds based on similarities between a feature vector of each of the unlabeled point clouds output through the model and feature vectors of the labeled point clouds output through the model and assigning a cluster label to each of the unlabeled point clouds based on a label of the selected labeled point cloud;
assigning pseudo labels to the unlabeled point clouds to which the cluster labels are assigned, wherein each pseudo label is assigned based on a confidence score obtained through the model;
updating the model by training the model with the labeled point clouds and the unlabeled point clouds to which the pseudo labels are assigned; and
performing selecting the labeled point cloud, assigning the pseudo labels, and updating the model by training iteratively after initially updating the model by training.
13. The automatic labeling method of claim 12 , wherein assigning the pseudo labels comprises:
selecting a part of the unlabeled point clouds based on silhouette scores thereof obtained through the model; and
assigning the pseudo labels, wherein each pseudo label is assigned based on a prediction result probability value obtained through the model for each point cloud of the selected part.
14. The automatic labeling method of claim 13 , wherein each pseudo label is assigned with a temporary label or the cluster label according to whether the temporary label and the cluster label are the same, and wherein the temporary label is determined based on the prediction result probability value obtained through the model.
15. The automatic labeling method of claim 14 , wherein the unlabeled point clouds to which pseudo labels are not assigned in the selected part remain unlabeled.
16. The automatic labeling method of claim 13 , wherein assigning the pseudo labels further comprises assigning the pseudo labels to the unlabeled point clouds that are not included in the selected part, wherein each pseudo label is assigned based on the prediction result probability value obtained through the model.
17. The automatic labeling method of claim 16 , wherein each pseudo label is assigned to the unlabeled point cloud if a temporary label corresponds to a class that is not in the feature vectors of the unlabeled point clouds of the selected part, and wherein the temporary label is determined based on the prediction result probability value obtained through the model for the corresponding unlabeled point cloud.
18. The automatic labeling method of claim 17 , wherein the unlabeled point clouds to which pseudo labels are not assigned among the unlabeled point clouds not included in the selected part remain unlabeled.
19. The automatic labeling method of claim 12 , further comprising terminating performing selecting the labeled point cloud, assigning the pseudo labels, and updating the model by training iteratively in response to the pseudo labels no longer being assigned.
20. The automatic labeling method of claim 12 , wherein selecting the labeled point cloud comprises:
selecting a preset number of labeled point clouds for each of the unlabeled point clouds based on the similarities; and
assigning, as the cluster label, a label with a highest frequency among labels of the selected labeled point clouds.